Currently, the heap arena map is a single, large array that covers
every possible arena frame in the entire address space. This is
practical up to about 48 bits of address space with 64 MB arenas.
However, there are two problems with this:
1. mips64, ppc64, and s390x support full 64-bit address spaces (though
on Linux only s390x has kernel support for 64-bit address spaces).
On these platforms, it would be good to support these larger
address spaces.
2. On Windows, processes are charged for untouched memory, so for
processes with small heaps, the mostly-untouched 32 MB arena map
plus a 64 MB arena are significant overhead. Hence, it would be
good to reduce both the arena map size and the arena size, but with
a single-level arena, these are inversely proportional.
This CL adds support for a two-level arena map. Arena frame numbers
are now divided into arenaL1Bits of L1 index and arenaL2Bits of L2
index.
At the moment, arenaL1Bits is always 0, so we effectively have a
single level map. We do a few things so that this has no cost beyond
the current single-level map:
1. We embed the L2 array directly in mheap, so if there's a single
entry in the L2 array, the representation is identical to the
current representation and there's no extra level of indirection.
2. Hot code that accesses the arena map is structured so that it
optimizes to nearly the same machine code as it does currently.
3. We make some small tweaks to hot code paths and to the inliner
itself to keep some important functions inlined despite their
now-larger ASTs. In particular, this is necessary for
heapBitsForAddr and heapBits.next.
Possibly as a result of some of the tweaks, this actually slightly
improves the performance of the x/benchmarks garbage benchmark:
name old time/op new time/op delta
Garbage/benchmem-MB=64-12 2.28ms ± 1% 2.26ms ± 1% -1.07% (p=0.000 n=17+19)
(https://perf.golang.org/search?q=upload:20180223.2)
For #23900.
Change-Id: If5164e0961754f97eb9eca58f837f36d759505ff
Reviewed-on: https://go-review.googlesource.com/96779
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Rick Hudson <rlh@golang.org>
There are too many places where I want to talk about "indexing into
the arena index". Make this less awkward and ambiguous by calling it
the "arena map" instead.
Change-Id: I726b0667bb2139dbc006175a0ec09a871cdf73f9
Reviewed-on: https://go-review.googlesource.com/96777
Run-TryBot: Austin Clements <austin@google.com>
Reviewed-by: Rick Hudson <rlh@golang.org>
Now that we support the full non-contiguous virtual address space of
amd64 hardware, some of the comments and constants related to this are
out of date.
This renames memLimitBits to heapAddrBits because 1<<memLimitBits is
no longer the limit of the address space and rewrites the comment to
focus first on hardware limits (which span OSes) and then discuss
kernel limits.
Second, this eliminates the memLimit constant because there's no
longer a meaningful "highest possible heap pointer value" on amd64.
Updates #23862.
Change-Id: I44b32033d2deb6b69248fb8dda14fc0e65c47f11
Reviewed-on: https://go-review.googlesource.com/95498
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Rick Hudson <rlh@golang.org>
On amd64, the virtual address space, when interpreted as signed
values, is [-2^47, 2^47). Currently, we only support heap addresses in
the "positive" half of this, [0, 2^47). This suffices for linux/amd64
and windows/amd64, but solaris/amd64 can map user addresses in the
negative part of this range. Specifically, addresses
0xFFFF8000'00000000 to 0xFFFFFD80'00000000 are part of user space.
This leads to "memory allocated by OS not in usable address space"
panic, since we don't map heap arena index space for these addresses.
Fix this by offsetting addresses when computing arena indexes so that
arena entry 0 corresponds to address -2^47 on amd64. We already map
enough arena space for 2^48 heap addresses on 64-bit (because arm64's
virtual address space is [0, 2^48)), so we don't need to grow any
structures to support this.
A different approach would be to simply mask out the top 16 bits.
However, there are two advantages to the offset approach: 1) invalid
heap addresses continue to naturally map to invalid arena indexes so
we don't need extra checks and 2) it perturbs the mapping of addresses
to arena indexes more, which helps check that we don't accidentally
compute incorrect arena indexes somewhere that happen to be right most
of the time.
Several comments and constant names are now somewhat misleading. We'll
fix that in the next CL. This CL is the core change the arena
indexing.
Fixes#23862.
Change-Id: Idb8e299fded04593a286b01a9582da6ddbac2f9a
Reviewed-on: https://go-review.googlesource.com/95497
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Rick Hudson <rlh@golang.org>
Accessing the arena index is about to get slightly more complicated.
Abstract this away into a set of functions for going back and forth
between addresses and arena slice indexes.
For #23862.
Change-Id: I0b20e74ef47a07b78ed0cf0a6128afe6f6e40f4b
Reviewed-on: https://go-review.googlesource.com/95496
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Rick Hudson <rlh@golang.org>
We don't want to account the memory for mheap_.arenas because most of
it is never touched, so currently we pass the address of a uint64 on
the heap. However, at least on mips, it's possible for this uint64 to
be unaligned, which causes the atomic add in mSysStatInc to crash.
Fix this by instead passing a nil stat pointer.
Fixes#23946.
Change-Id: I091587df1b3066c330b6bb4d834e4596c407910f
Reviewed-on: https://go-review.googlesource.com/95695
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Cherry Zhang <cherryyz@google.com>
reflect.unsafe_New is an often called function according
to profiling in a large production environment.
Since newobject is not inlined currently there
is call overhead that can be avoided by calling
mallocgc directly.
name old time/op new time/op delta
New 32.4ns ± 2% 29.8ns ± 1% -8.03% (p=0.000 n=19+20)
Change-Id: I572e4be830ed8e5c0da555dc3a8864c8363112be
Reviewed-on: https://go-review.googlesource.com/95015
Reviewed-by: Austin Clements <austin@google.com>
Now that we have memLimit, also having _MaxMem is a bit confusing.
Replace it with maxAlloc, which better conveys what it limits. We also
define maxAlloc slightly differently: since it's now clear that it
limits allocation size, we can account for a subtle difference between
32-bit and 64-bit.
Change-Id: Iac39048018cc0dae7f0919e25185fee4b3eed529
Reviewed-on: https://go-review.googlesource.com/85890
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Rick Hudson <rlh@golang.org>
Currently there's a detailed comment in lfstack_64bit.go about address
space limitations on various architectures. Since that's now relevant
to malloc, move it to a more prominent place in the documentation for
memLimitBits.
Updates #10460.
Change-Id: If9708291cf3a288057b8b3ba0ba6a59e3602bbd6
Reviewed-on: https://go-review.googlesource.com/85889
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Rick Hudson <rlh@golang.org>
Currently large sysReserve calls on some OSes don't actually reserve
the memory, but just check that it can be reserved. This was important
when we called sysReserve to "reserve" many gigabytes for the heap up
front, but now that we map memory in small increments as we need it,
this complication is no longer necessary.
This has one curious side benefit: currently, on Linux, allocations
that are large enough to be rejected by mmap wind up freezing the
application for a long time before it panics. This happens because
sysReserve doesn't reserve the memory, so sysMap calls mmap_fixed,
which calls mmap, which fails because the mapping is too large.
However, mmap_fixed doesn't inspect *why* mmap fails, so it falls back
to probing every page in the desired region individually with mincore
before performing an (otherwise dangerous) MAP_FIXED mapping, which
will also fail. This takes a long time for a large region. Now this
logic is gone, so the mmap failure leads to an immediate panic.
Updates #10460.
Change-Id: I8efe88c611871cdb14f99fadd09db83e0161ca2e
Reviewed-on: https://go-review.googlesource.com/85888
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Rick Hudson <rlh@golang.org>
This replaces the contiguous heap arena mapping with a potentially
sparse mapping that can support heap mappings anywhere in the address
space.
This has several advantages over the current approach:
* There is no longer any limit on the size of the Go heap. (Currently
it's limited to 512GB.) Hence, this fixes#10460.
* It eliminates many failures modes of heap initialization and
growing. In particular it eliminates any possibility of panicking
with an address space conflict. This can happen for many reasons and
even causes a low but steady rate of TSAN test failures because of
conflicts with the TSAN runtime. See #16936 and #11993.
* It eliminates the notion of "non-reserved" heap, which was added
because creating huge address space reservations (particularly on
64-bit) led to huge process VSIZE. This was at best confusing and at
worst conflicted badly with ulimit -v. However, the non-reserved
heap logic is complicated, can race with other mappings in non-pure
Go binaries (e.g., #18976), and requires that the entire heap be
either reserved or non-reserved. We currently maintain the latter
property, but it's quite difficult to convince yourself of that, and
hence difficult to keep correct. This logic is still present, but
will be removed in the next CL.
* It fixes problems on 32-bit where skipping over parts of the address
space leads to mapping huge (and never-to-be-used) metadata
structures. See #19831.
This also completely rewrites and significantly simplifies
mheap.sysAlloc, which has been a source of many bugs. E.g., #21044,
#20259, #18651, and #13143 (and maybe #23222).
This change also makes it possible to allocate individual objects
larger than 512GB. As a result, a few tests that expected huge
allocations to fail needed to be changed to make even larger
allocations. However, at the moment attempting to allocate a humongous
object may cause the program to freeze for several minutes on Linux as
we fall back to probing every page with addrspace_free. That logic
(and this failure mode) will be removed in the next CL.
Fixes#10460.
Fixes#22204 (since it rewrites the code involved).
This slightly slows down compilebench and the x/benchmarks garbage
benchmark.
name old time/op new time/op delta
Template 184ms ± 1% 185ms ± 1% ~ (p=0.065 n=10+9)
Unicode 86.9ms ± 3% 86.3ms ± 1% ~ (p=0.631 n=10+10)
GoTypes 599ms ± 0% 602ms ± 0% +0.56% (p=0.000 n=10+9)
Compiler 2.87s ± 1% 2.89s ± 1% +0.51% (p=0.002 n=9+10)
SSA 7.29s ± 1% 7.25s ± 1% ~ (p=0.182 n=10+9)
Flate 118ms ± 2% 118ms ± 1% ~ (p=0.113 n=9+9)
GoParser 147ms ± 1% 148ms ± 1% +1.07% (p=0.003 n=9+10)
Reflect 401ms ± 1% 404ms ± 1% +0.71% (p=0.003 n=10+9)
Tar 175ms ± 1% 175ms ± 1% ~ (p=0.604 n=9+10)
XML 209ms ± 1% 210ms ± 1% ~ (p=0.052 n=10+10)
(https://perf.golang.org/search?q=upload:20171231.4)
name old time/op new time/op delta
Garbage/benchmem-MB=64-12 2.23ms ± 1% 2.25ms ± 1% +0.84% (p=0.000 n=19+19)
(https://perf.golang.org/search?q=upload:20171231.3)
Relative to the start of the sparse heap changes (starting at and
including "runtime: fix various contiguous bitmap assumptions"),
overall slowdown is roughly 1% on GC-intensive benchmarks:
name old time/op new time/op delta
Template 183ms ± 1% 185ms ± 1% +1.32% (p=0.000 n=9+9)
Unicode 84.9ms ± 2% 86.3ms ± 1% +1.65% (p=0.000 n=9+10)
GoTypes 595ms ± 1% 602ms ± 0% +1.19% (p=0.000 n=9+9)
Compiler 2.86s ± 0% 2.89s ± 1% +0.91% (p=0.000 n=9+10)
SSA 7.19s ± 0% 7.25s ± 1% +0.75% (p=0.000 n=8+9)
Flate 117ms ± 1% 118ms ± 1% +1.10% (p=0.000 n=10+9)
GoParser 146ms ± 2% 148ms ± 1% +1.48% (p=0.002 n=10+10)
Reflect 398ms ± 1% 404ms ± 1% +1.51% (p=0.000 n=10+9)
Tar 173ms ± 1% 175ms ± 1% +1.17% (p=0.000 n=10+10)
XML 208ms ± 1% 210ms ± 1% +0.62% (p=0.011 n=10+10)
[Geo mean] 369ms 373ms +1.17%
(https://perf.golang.org/search?q=upload:20180101.2)
name old time/op new time/op delta
Garbage/benchmem-MB=64-12 2.22ms ± 1% 2.25ms ± 1% +1.51% (p=0.000 n=20+19)
(https://perf.golang.org/search?q=upload:20180101.3)
Change-Id: I5daf4cfec24b252e5a57001f0a6c03f22479d0f0
Reviewed-on: https://go-review.googlesource.com/85887
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Rick Hudson <rlh@golang.org>
This replaces all uses of the mheap_.arena_* fields outside of
mallocinit and sysAlloc. These fields fundamentally assume a
contiguous heap between two bounds, so eliminating these is necessary
for a sparse heap.
Many of these are replaced with checks for non-nil spans at the test
address (which in turn checks for a non-nil entry in the heap arena
array). Some of them are just for debugging and somewhat meaningless
with a sparse heap, so those we just delete.
Updates #10460.
Change-Id: I8345b95ffc610aed694f08f74633b3c63506a41f
Reviewed-on: https://go-review.googlesource.com/85886
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Rick Hudson <rlh@golang.org>
This splits the heap bitmap into separate chunks for every 64MB of the
heap and introduces an index mapping from virtual address to metadata.
It modifies the heapBits abstraction to use this two-level structure.
Finally, it modifies heapBitsSetType to unroll the bitmap into the
object itself and then copy it out if the bitmap would span
discontiguous bitmap chunks.
This is a step toward supporting general sparse heaps, which will
eliminate address space conflict failures as well as the limit on the
heap size.
It's also advantageous for 32-bit. 32-bit already supports
discontiguous heaps by always starting the arena at address 0.
However, as a result, with a contiguous bitmap, if the kernel chooses
a high address (near 2GB) for a heap mapping, the runtime is forced to
map up to 128MB of heap bitmap. Now the runtime can map sections of
the bitmap for just the parts of the address space used by the heap.
Updates #10460.
This slightly slows down the x/garbage and compilebench benchmarks.
However, I think the slowdown is acceptably small.
name old time/op new time/op delta
Template 178ms ± 1% 180ms ± 1% +0.78% (p=0.029 n=10+10)
Unicode 85.7ms ± 2% 86.5ms ± 2% ~ (p=0.089 n=10+10)
GoTypes 594ms ± 0% 599ms ± 1% +0.70% (p=0.000 n=9+9)
Compiler 2.86s ± 0% 2.87s ± 0% +0.40% (p=0.001 n=9+9)
SSA 7.23s ± 2% 7.29s ± 2% +0.94% (p=0.029 n=10+10)
Flate 116ms ± 1% 117ms ± 1% +0.99% (p=0.000 n=9+9)
GoParser 146ms ± 1% 146ms ± 0% ~ (p=0.193 n=10+7)
Reflect 399ms ± 0% 403ms ± 1% +0.89% (p=0.001 n=10+10)
Tar 173ms ± 1% 174ms ± 1% +0.91% (p=0.013 n=10+9)
XML 208ms ± 1% 210ms ± 1% +0.93% (p=0.000 n=10+10)
[Geo mean] 368ms 371ms +0.79%
name old time/op new time/op delta
Garbage/benchmem-MB=64-12 2.17ms ± 1% 2.21ms ± 1% +2.15% (p=0.000 n=20+20)
Change-Id: I037fd283221976f4f61249119d6b97b100bcbc66
Reviewed-on: https://go-review.googlesource.com/85883
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Rick Hudson <rlh@golang.org>
Currently the heap bitamp is laid in reverse order in memory relative
to the heap itself. This was originally done out of "excessive
cleverness" so that computing a bitmap pointer could load only the
arena_start field and so that heaps could be more contiguous by
growing the arena and the bitmap out from a common center point.
However, this appears to have no actual performance benefit, it
complicates nearly every use of the bitmap, and it makes already
confusing code more confusing. Furthermore, it's still possible to use
a single field (the new bitmap_delta) for the bitmap pointer
computation by employing slightly different excessive cleverness.
Hence, this CL puts the bitmap into forward order.
This is a (very) updated version of CL 9404.
Change-Id: I743587cc626c4ecd81e660658bad85b54584108c
Reviewed-on: https://go-review.googlesource.com/85881
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Rick Hudson <rlh@golang.org>
We're about to start tracking nowritebarrierrec through systemstack
calls, which will reveal write barriers in persistentalloc prohibited
by various callers.
The pointers manipulated by persistentalloc are always to off-heap
memory, so this removes these write barriers statically by introducing
a new go:notinheap type to represent generic off-heap memory.
Updates #22384.
For #22460.
Change-Id: Id449d9ebf145b14d55476a833e7f076b0d261d57
Reviewed-on: https://go-review.googlesource.com/72771
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Rick Hudson <rlh@golang.org>
The previous comment of nextSample didn't mention Poisson processes,
which is the reason why it needed to create an exponential
distribution, so it was hard to follow the reasoning for people
not highly familiar with statistics.
Since we're at it, we also make it clear that we are just creating
a random number with exponential distribution by moving the
bulk of the function into a new fastexprand().
No functional changes.
Change-Id: I9c275e87edb3418ee0974257af64c73465028ad7
Reviewed-on: https://go-review.googlesource.com/65657
Reviewed-by: Austin Clements <austin@google.com>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
https://golang.org/cl/22598 made nextFreeFast inlinable.
But during https://golang.org/cl/63611 it was discovered, that it is no longer inlinable.
Reduce number of statements below inlining threshold to make it inlinable again.
Also update tests, to prevent regressions.
Doesn't reduce readability.
Change-Id: Ia672784dd48ed3b1ab46e390132f1094fe453de5
Reviewed-on: https://go-review.googlesource.com/65030
Run-TryBot: Ilya Tocar <ilya.tocar@intel.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Daniel Martí <mvdan@mvdan.cc>
This avoids division and multiplication.
Instrumentation suggests that this is a very common case.
Change-Id: I2d5d5012d4f4df4c4af1f9f85ca9c323c9889c0e
Reviewed-on: https://go-review.googlesource.com/54657
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
Right now, if it's possible to grow the arena reservation but
mheap.sysAlloc fails to get 256MB more of memory, it simply fails.
However, on 32-bit we have a fallback path that uses much smaller
mmaps that could take in this situation, but fail to.
This commit fixes mheap.sysAlloc to use a common failure path in case
it can't grow the reservation. On 32-bit, this path includes the
fallback.
Ideally, mheap.sysAlloc would attempt smaller reservation growths
first, but taking the fallback path is a simple change for Go 1.9.
Updates #21044 (fixes one of two issues).
Change-Id: I1e0035ffba986c3551479d5742809e43da5e7c73
Reviewed-on: https://go-review.googlesource.com/51713
Run-TryBot: Austin Clements <austin@google.com>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Currently, the heap arena allocator allocates monotonically increasing
addresses. This is fine on 64-bit where we stake out a giant block of
the address space for ourselves and start at the beginning of it, but
on 32-bit the arena starts at address 0 but we start allocating from
wherever the OS feels like giving us memory. We can generally hint the
OS to start us at a low address, but this doesn't always work.
As a result, on 32-bit, if the OS gives us an arena block that's lower
than the current block we're allocating from, we simply say "thanks
but no thanks", return the whole (256MB!) block of memory, and then
take a fallback path that mmaps just the amount of memory we need
(which may be as little as 8K).
We have to do this because mheap_.arena_used is *both* the highest
used address in the arena and the next address we allocate from.
Fix all of this by separating the second role of arena_used out into a
new field called arena_alloc. This lets us accept any arena block the
OS gives us. This also slightly changes the invariants around
arena_end. Previously, we ensured arena_used <= arena_end, but this
was related to arena_used's second role, so the new invariant is
arena_alloc <= arena_end. As a result, we no longer necessarily update
arena_end when we're updating arena_used.
Fixes#20259 properly. (Unlike the original fix, this one should not
be cherry-picked to Go 1.8.)
This is reasonably low risk. I verified several key properties of the
32-bit code path with both 4K and 64K physical pages using a symbolic
model and the change does not materially affect 64-bit (arena_used ==
arena_alloc on 64-bit). The only oddity is that we no longer call
setArenaUsed with racemap == false to indicate that we're creating a
hole in the address space, but this only happened in a 32-bit-only
code path, and the race detector require 64-bit, so this never
mattered anyway.
Change-Id: Ib1334007933e615166bac4159bf357ae06ec6a25
Reviewed-on: https://go-review.googlesource.com/44010
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
Reviewed-by: Rick Hudson <rlh@golang.org>
If mheap.sysAlloc doesn't have room in the heap arena for an
allocation, it will attempt to map more address space with sysReserve.
sysReserve is given a hint, but can return any unused address range.
Currently, mheap.sysAlloc incorrectly assumes the returned region will
never fall between arena_start and arena_used. If it does,
mheap.sysAlloc will blindly accept the new region as the new
arena_used and arena_end, causing these to decrease and make it so any
Go heap above the new arena_used is no longer considered part of the
Go heap. This assumption *used to be* safe because we had all memory
between arena_start and arena_used mapped, but when we switched to an
arena_start of 0 on 32-bit, it became no longer safe.
Most likely, we've only recently seen this bug occur because we
usually start arena_used just above the binary, which is low in the
address space. Hence, the kernel is very unlikely to give us a region
before arena_used.
Since mheap.sysAlloc is a linear allocator, there's not much we can do
to handle this well. Hence, we fix this problem by simply rejecting
the new region if it isn't after arena_end. In this case, we'll take
the fall-back path and mmap a small region at any address just for the
requested memory.
Fixes#20259.
Change-Id: Ib72e8cd621545002d595c7cade1e817cfe3e5b1e
Reviewed-on: https://go-review.googlesource.com/43870
Reviewed-by: Ian Lance Taylor <iant@golang.org>
Reviewed-by: Rick Hudson <rlh@golang.org>
Currently _TinySizeClass is untyped, which means it can accidentally
be used as a spanClass (not that I would know this from experience or
anything). Make it an int8 to avoid this mix up.
This is a cherry-pick of dev.garbage commit 81b74bf9c5.
Change-Id: I1e69eccee436ea5aa45e9a9828a013e369e03f1a
Reviewed-on: https://go-review.googlesource.com/41254
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Rick Hudson <rlh@golang.org>
Currently, we mix objects with pointers and objects without pointers
("noscan" objects) together in memory. As a result, for every object
we grey, we have to check that object's heap bits to find out if it's
noscan, which adds to the per-object cost of GC. This also hurts the
TLB footprint of the garbage collector because it decreases the
density of scannable objects at the page level.
This commit improves the situation by using separate spans for noscan
objects. This will allow a much simpler noscan check (in a follow up
CL), eliminate the need to clear the bitmap of noscan objects (in a
follow up CL), and improves TLB footprint by increasing the density of
scannable objects.
This is also a step toward eliminating dead bits, since the current
noscan check depends on checking the dead bit of the first word.
This has no effect on the heap size of the garbage benchmark.
We'll measure the performance change of this after the follow-up
optimizations.
This is a cherry-pick from dev.garbage commit d491e550c3. The only
non-trivial merge conflict was in updatememstats in mstats.go, where
we now have to separate the per-spanclass stats from the per-sizeclass
stats.
Change-Id: I13bdc4869538ece5649a8d2a41c6605371618e40
Reviewed-on: https://go-review.googlesource.com/41251
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Rick Hudson <rlh@golang.org>
On 32-bit architectures (or if we fail to map a 64-bit-style arena),
we try to map the heap arena just above the end of the process image.
While we can accept any address, using lower addresses is preferable
because lower addresses cause us to map less of the heap bitmap.
However, if a program is linked against C code that has global
constructors, those constructors may call brk/sbrk to allocate memory
(e.g., many C malloc implementations do this for small allocations).
The brk also starts just above the process image, so this may adjust
the brk past the beginning of where we want to put the heap arena. In
this case, the kernel will pick a different address for the arena and
it will usually be very high (at least, as these things go in a 32-bit
address space).
Fix this by consulting the current value of the brk and using this in
addition to the end of the process image to compute the initial arena
placement.
This is implemented only on Linux currently, since we have no evidence
that it's an issue on any other OSes.
Fixes#19831.
Change-Id: Id64b45d08d8c91e4f50d92d0339146250b04f2f8
Reviewed-on: https://go-review.googlesource.com/39810
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
Changing mheap_.arena_used requires several steps that are currently
repeated multiple times in mheap_.sysAlloc. Consolidate these into a
single function.
In the future, this will also make it easier to add other auxiliary VM
structures.
Change-Id: Ie68837d2612e1f4ba4904acb1b6b832b15431d56
Reviewed-on: https://go-review.googlesource.com/40151
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
Currently the GC triggering condition is an awkward combination of the
gcMode (whether or not it's gcBackgroundMode) and a boolean
"forceTrigger" flag.
Replace this with a new gcTrigger type that represents the range of
transition predicates we need. This has several advantages:
1. We can remove the awkward logic that affects the trigger behavior
based on the gcMode. Now gcMode purely controls whether to run a
STW GC or not and the gcTrigger controls whether this is a forced
GC that cannot be consolidated with other GC cycles.
2. We can lift the time-based triggering logic in sysmon to just
another type of GC trigger and move the logic to the trigger test.
3. This sets us up to have a cycle count-based trigger, which we'll
use to make runtime.GC trigger concurrent GC with the desired
consolidation properties.
For #18216.
Change-Id: If9cd49349579a548800f5022ae47b8128004bbfc
Reviewed-on: https://go-review.googlesource.com/37516
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Rick Hudson <rlh@golang.org>
Implement math/bits.TrailingZerosX using intrinsics.
Generally reorganize the intrinsic spec a bit.
The instrinsics data structure is now built at init time.
This will make doing the other functions in math/bits easier.
Update sys.CtzX to return int instead of uint{64,32} so it
matches math/bits.TrailingZerosX.
Improve the intrinsics a bit for amd64. We don't need the CMOV
for <64 bit versions.
Update #18616
Change-Id: Ic1c5339c943f961d830ae56f12674d7b29d4ff39
Reviewed-on: https://go-review.googlesource.com/38155
Run-TryBot: Keith Randall <khr@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
Extend period of fastrand from (1<<31)-1 to (1<<32)-1 by
choosing other polynom and reacting on high bit before shift.
Polynomial is taken at https://users.ece.cmu.edu/~koopman/lfsr/index.html
from 32.dat.gz . It is referred as F7711115 cause this list of
polynomials is for LFSR with shift to right (and fastrand uses shift to
left). (old polynomial is referred in 31.dat.gz as 7BB88888).
There were couple of places with conversation of fastrand to int, which
leads to negative values on 32bit platforms. They are fixed.
Change-Id: Ibee518a3f9103e0aea220ada494b3aec77babb72
Reviewed-on: https://go-review.googlesource.com/36875
Run-TryBot: Minux Ma <minux@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Minux Ma <minux@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
Currently both _MaxMem and _MaxArena32 represent the maximum arena
size on 32-bit hosts (except on MIPS32 where _MaxMem is confusingly
smaller than _MaxArena32).
Clean up sysAlloc so that it always uses _MaxMem, which is the maximum
arena size on both 32- and 64-bit architectures and is the arena size
we allocate auxiliary structures for. This lets us simplify and unify
some code paths and eliminate _MaxArena32.
Fixes#18651. mheap.sysAlloc currently assumes that if the arena is
small, we must be on a 32-bit machine and can therefore grow the arena
to _MaxArena32. This breaks down on darwin/arm64, where _MaxMem is
only 2 GB. As a result, on darwin/arm64, we only reserve spans and
bitmap space for a 2 GB heap, and if the application tries to allocate
beyond that, sysAlloc takes the 32-bit path, tries to grow the arena
beyond 2 GB, and panics when it tries to grow the spans array
allocation past its reserved size. This has probably been a problem
for several releases now, but was only noticed recently because
mapSpans didn't check the bounds on the span reservation until
recently. Most likely it corrupted the bitmap before. By using _MaxMem
consistently, we avoid thinking that we can grow the arena larger than
we have auxiliary structures for.
Change-Id: Ifef28cb746a3ead4b31c1d7348495c2242fef520
Reviewed-on: https://go-review.googlesource.com/35253
Reviewed-by: David Crawshaw <crawshaw@golang.org>
Reviewed-by: Elias Naur <elias.naur@gmail.com>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
mallocinit has evolved organically. Make a pass to clean it up in
various ways:
1. Merge the computation of spansSize and bitmapSize. These were
computed on every loop iteration of two different loops, but always
have the same value, which can be derived directly from _MaxMem.
This also avoids over-reserving these on MIPS, were _MaxArena32 is
larger than _MaxMem.
2. Remove the ulimit -v logic. It's been disabled for many releases
and the dead code paths to support it are even more wrong now than
they were when it was first disabled, since now we *must* reserve
spans and bitmaps for the full address space.
3. Make it clear that we're using a simple linear allocation to lay
out the spans, bitmap, and arena spaces. Previously there were a
lot of redundant pointer computations. Now we just bump p1 up as we
reserve the spaces.
In preparation for #18651.
Updates #5049 (respect ulimit).
Change-Id: Icbe66570d3a7a17bea227dc54fb3c4978b52a3af
Reviewed-on: https://go-review.googlesource.com/35252
Reviewed-by: Russ Cox <rsc@golang.org>
Currently _MaxMem is a uintptr, which is going to complicate some
further changes. Make it untyped so we'll be able to do untyped math
on it before truncating it to a uintptr.
The runtime assembly is identical before and after this change on
{linux,windows}/{amd64,386}.
Updates #18651.
Change-Id: I0f64511faa9e0aa25179a556ab9f185ebf8c9cf8
Reviewed-on: https://go-review.googlesource.com/35251
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Rick Hudson <rlh@golang.org>
Reviewed-by: David Crawshaw <crawshaw@golang.org>
Currently, the check for legal pointers in stack copying uses
_PageSize (8K) as the minimum legal pointer. By default, Linux won't
let you map under 64K, but
1) it's less clear what other OSes allow or will allow in the future;
2) while mapping the first page is a terrible idea, mapping anywhere
above that is arguably more justifiable;
3) the compiler only assumes the first physical page (4K) is never
mapped.
Make the runtime consistent with the compiler and more robust by
changing the bad pointer check to use 4K as the minimum legal pointer.
This came out of discussions on CLs 34663 and 34719.
Change-Id: Idf721a788bd9699fb348f47bdd083cf8fa8bd3e5
Reviewed-on: https://go-review.googlesource.com/34890
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Russ Cox <rsc@golang.org>
No point in computing this info on startup.
Compute it at build time.
This lets us spend more time computing & checking the size classes.
Improve the div magic for rounding to the start of an object.
We can now use 32-bit multiplies & shifts, which should help
32-bit platforms.
The static data is <1KB.
The actual size classes are not changed by this CL.
Change-Id: I6450cec7d1b2b4ad31fd3f945f504ed2ec6570e7
Reviewed-on: https://go-review.googlesource.com/32219
Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Austin Clements <austin@google.com>
Since barrier-less memclr is only safe in very narrow circumstances,
this commit renames memclr to avoid accidentally calling memclr on
typed memory. This can cause subtle, non-deterministic bugs, so it's
worth some effort to prevent. In the near term, this will also prevent
bugs creeping in from any concurrent CLs that add calls to memclr; if
this happens, whichever patch hits master second will fail to compile.
This also adds the other new memclr variants to the compiler's
builtin.go to minimize the churn on that binary blob. We'll use these
in future commits.
Updates #17503.
Change-Id: I00eead049f5bd35ca107ea525966831f3d1ed9ca
Reviewed-on: https://go-review.googlesource.com/31369
Reviewed-by: Keith Randall <khr@golang.org>
Reviewed-by: Rick Hudson <rlh@golang.org>
Currently fixalloc does not zero memory it reuses. This is dangerous
with the hybrid barrier if the type may contain heap pointers, since
it may cause us to observe a dead heap pointer on reuse. It's also
error-prone since it's the only allocator that doesn't zero on
allocation (mallocgc of course zeroes, but so do persistentalloc and
sysAlloc). It's also largely pointless: for mcache, the caller
immediately memclrs the allocation; and the two specials types are
tiny so there's no real cost to zeroing them.
Change fixalloc to zero allocations by default.
The only type we don't zero by default is mspan. This actually
requires that the spsn's sweepgen survive across freeing and
reallocating a span. If we were to zero it, the following race would
be possible:
1. The current sweepgen is 2. Span s is on the unswept list.
2. Direct sweeping sweeps span s, finds it's all free, and releases s
to the fixalloc.
3. Thread 1 allocates s from fixalloc. Suppose this zeros s, including
s.sweepgen.
4. Thread 1 calls s.init, which sets s.state to _MSpanDead.
5. On thread 2, background sweeping comes across span s in allspans
and cas's s.sweepgen from 0 (sg-2) to 1 (sg-1). Now it thinks it
owns it for sweeping. 6. Thread 1 continues initializing s.
Everything breaks.
I would like to fix this because it's obviously confusing, but it's a
subtle enough problem that I'm leaving it alone for now. The solution
may be to skip sweepgen 0, but then we have to think about wrap-around
much more carefully.
Updates #17503.
Change-Id: Ie08691feed3abbb06a31381b94beb0a2e36a0613
Reviewed-on: https://go-review.googlesource.com/31368
Reviewed-by: Keith Randall <khr@golang.org>
Reviewed-by: Rick Hudson <rlh@golang.org>
Like h_allspans and mheap_.allspans, these were two ways of referring
to the spans array from when the runtime was split between C and Go.
Clean this up by making mheap_.spans a slice and eliminating h_spans.
Change-Id: I3aa7038d53c3a4252050aa33e468c48dfed0b70e
Reviewed-on: https://go-review.googlesource.com/30532
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Rick Hudson <rlh@golang.org>
This covers basically all sysAlloc'd, persistentalloc'd, and
fixalloc'd types.
Change-Id: I0487c887c2a0ade5e33d4c4c12d837e97468e66b
Reviewed-on: https://go-review.googlesource.com/30941
Reviewed-by: Rick Hudson <rlh@golang.org>
The big documentation comment at the top of malloc.go has gotten
woefully out of date. Update it.
Change-Id: Ibdb1bdcfdd707a6dc9db79d0633a36a28882301b
Reviewed-on: https://go-review.googlesource.com/29731
Reviewed-by: Hyang-Ah Hana Kim <hyangah@gmail.com>
Reviewed-by: Rick Hudson <rlh@golang.org>
Now that the runtime fetches the true physical page size from the OS,
make the physical page size used by heap growth a variable instead of
a constant. This isn't used in any performance-critical paths, so it
shouldn't be an issue.
sys.PhysPageSize is also renamed to sys.DefaultPhysPageSize to make it
clear that it's not necessarily the true page size. There are no uses
of this constant any more, but we'll keep it around for now.
Updates #12480 and #10180.
Change-Id: I6c23b9df860db309c38c8287a703c53817754f03
Reviewed-on: https://go-review.googlesource.com/25022
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Rick Hudson <rlh@golang.org>
Currently the physical page size assumed by the runtime is hard-coded.
On Linux the runtime at least fetches the OS page size during init and
sanity checks against the hard-coded value, but they may still differ.
On other OSes we wouldn't even notice.
Add support on all OSes to fetch the actual OS physical page size
during runtime init and lift the sanity check of PhysPageSize from the
Linux init code to general malloc init. Currently this is the only use
of the retrieved page size, but we'll add more shortly.
Updates #12480 and #10180.
Change-Id: I065f2834bc97c71d3208edc17fd990ec9058b6da
Reviewed-on: https://go-review.googlesource.com/25050
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Rick Hudson <rlh@golang.org>
Currently we only execute a publication barrier for scan objects (and
skip it for noscan objects). This used to be okay because GC would
never consult the object itself (so it wouldn't observe uninitialized
memory even if it found a pointer to a noscan object), and the heap
bitmap was pre-initialized to noscan.
However, now we explicitly initialize the heap bitmap for noscan
objects when we allocate them. While the GC will still never consult
the contents of a noscan object, it does need to see the initialized
heap bitmap. Hence, we need to execute a publication barrier to make
the bitmap visible before user code can expose a pointer to the newly
allocated object even for noscan objects.
Change-Id: Ie4133c638db0d9055b4f7a8061a634d970627153
Reviewed-on: https://go-review.googlesource.com/23043
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Rick Hudson <rlh@golang.org>
In issue #13992, Russ mentioned that the heap bitmap footprint was
halved but that the bitmap size calculation hadn't been updated. This
presents the opportunity to either halve the bitmap size or double
the addressable virtual space. This CL doubles the addressable virtual
space. On 32 bit this can be tweaked further to allow the bitmap to
cover the entire 4GB virtual address space, removing a failure mode
if the kernel hands out memory with a too low address.
First, fix the calculation and double _MaxArena32 to cover 4GB virtual
memory space with the same bitmap size (256 MB).
Then, allow the fallback mode for the initial memory reservation
on 32 bit (or 64 bit with too little available virtual memory) to not
include space for the arena. mheap.sysAlloc will automatically reserve
additional space when the existing arena is full.
Finally, set arena_start to 0 in 32 bit mode, so that any address is
acceptable for subsequent (additional) reservations.
Before, the bitmap was always located just before arena_start, so
fix the two places relying on that assumption: Point the otherwise unused
mheap.bitmap to one byte after the end of the bitmap, and use it for
bitmap addressing instead of arena_start.
With arena_start set to 0 on 32 bit, the cgoInRange check is no longer a
sufficient check for Go pointers. Introduce and call inHeapOrStack to
check whether a pointer is to the Go heap or stack.
While we're here, remove sysReserveHigh which seems to be unused.
Fixes#13992
Change-Id: I592b513148a50b9d3967b5c5d94b86b3ec39acc2
Reviewed-on: https://go-review.googlesource.com/20471
Reviewed-by: Austin Clements <austin@google.com>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
With the switch to separate mark bitmaps, the scan/dead bit for the
first word of each object is now unused. Reclaim this bit and use it
as a scan/dead bit, just like words three and on. The second word is
still used for checkmark.
This dramatically simplifies heapBitsSetTypeNoScan and hasPointers,
since they no longer need different cases for 1, 2, and 3+ word
objects. They can instead just manipulate the heap bitmap for the
first word and be done with it.
In order to enable this, we change heapBitsSetType and runGCProg to
always set the scan/dead bit to scan for the first word on every code
path. Since these functions only apply to types that have pointers,
there's no need to do this conditionally: it's *always* necessary to
set the scan bit in the first word.
We also change every place that scans an object and checks if there
are more pointers. Rather than only checking morePointers if the word
is >= 2, we now check morePointers if word != 1 (since that's the
checkmark word).
Looking forward, we should probably reclaim the checkmark bit, too,
but that's going to be quite a bit more work.
Tested by setting doubleCheck in heapBitsSetType and running all.bash
on both linux/amd64 and linux/386, and by running GOGC=10 all.bash.
This particularly improves the FmtFprintf* go1 benchmarks, since they
do a large amount of noscan allocation.
name old time/op new time/op delta
BinaryTree17-12 2.34s ± 1% 2.38s ± 1% +1.70% (p=0.000 n=17+19)
Fannkuch11-12 2.09s ± 0% 2.09s ± 1% ~ (p=0.276 n=17+16)
FmtFprintfEmpty-12 44.9ns ± 2% 44.8ns ± 2% ~ (p=0.340 n=19+18)
FmtFprintfString-12 127ns ± 0% 125ns ± 0% -1.57% (p=0.000 n=16+15)
FmtFprintfInt-12 128ns ± 0% 122ns ± 1% -4.45% (p=0.000 n=15+20)
FmtFprintfIntInt-12 207ns ± 1% 193ns ± 0% -6.55% (p=0.000 n=19+14)
FmtFprintfPrefixedInt-12 197ns ± 1% 191ns ± 0% -2.93% (p=0.000 n=17+18)
FmtFprintfFloat-12 263ns ± 0% 248ns ± 1% -5.88% (p=0.000 n=15+19)
FmtManyArgs-12 794ns ± 0% 779ns ± 1% -1.90% (p=0.000 n=18+18)
GobDecode-12 7.14ms ± 2% 7.11ms ± 1% ~ (p=0.072 n=20+20)
GobEncode-12 5.85ms ± 1% 5.82ms ± 1% -0.49% (p=0.000 n=20+20)
Gzip-12 218ms ± 1% 215ms ± 1% -1.22% (p=0.000 n=19+19)
Gunzip-12 36.8ms ± 0% 36.7ms ± 0% -0.18% (p=0.006 n=18+20)
HTTPClientServer-12 77.1µs ± 4% 77.1µs ± 3% ~ (p=0.945 n=19+20)
JSONEncode-12 15.6ms ± 1% 15.9ms ± 1% +1.68% (p=0.000 n=18+20)
JSONDecode-12 55.2ms ± 1% 53.6ms ± 1% -2.93% (p=0.000 n=17+19)
Mandelbrot200-12 4.05ms ± 1% 4.05ms ± 0% ~ (p=0.306 n=17+17)
GoParse-12 3.14ms ± 1% 3.10ms ± 1% -1.31% (p=0.000 n=19+18)
RegexpMatchEasy0_32-12 69.3ns ± 1% 70.0ns ± 0% +0.89% (p=0.000 n=19+17)
RegexpMatchEasy0_1K-12 237ns ± 1% 236ns ± 0% -0.62% (p=0.000 n=19+16)
RegexpMatchEasy1_32-12 69.5ns ± 1% 70.3ns ± 1% +1.14% (p=0.000 n=18+17)
RegexpMatchEasy1_1K-12 377ns ± 1% 366ns ± 1% -3.03% (p=0.000 n=15+19)
RegexpMatchMedium_32-12 107ns ± 1% 107ns ± 2% ~ (p=0.318 n=20+19)
RegexpMatchMedium_1K-12 33.8µs ± 3% 33.5µs ± 1% -1.04% (p=0.001 n=20+19)
RegexpMatchHard_32-12 1.68µs ± 1% 1.73µs ± 0% +2.50% (p=0.000 n=20+18)
RegexpMatchHard_1K-12 50.8µs ± 1% 52.0µs ± 1% +2.50% (p=0.000 n=19+18)
Revcomp-12 381ms ± 1% 385ms ± 1% +1.00% (p=0.000 n=17+18)
Template-12 64.9ms ± 3% 62.6ms ± 1% -3.55% (p=0.000 n=19+18)
TimeParse-12 324ns ± 0% 328ns ± 1% +1.25% (p=0.000 n=18+18)
TimeFormat-12 345ns ± 0% 334ns ± 0% -3.31% (p=0.000 n=15+17)
[Geo mean] 52.1µs 51.5µs -1.00%
Change-Id: I13e74da3193a7f80794c654f944d1f0d60817049
Reviewed-on: https://go-review.googlesource.com/22632
Reviewed-by: Rick Hudson <rlh@golang.org>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
nextFreeFast is currently not inlined by the compiler due
to its size and complexity. This CL simplifies
nextFreeFast by letting the slow path handle (nextFree)
handle a corner cases.
Change-Id: Ia9c5d1a7912bcb4bec072f5fd240f0e0bafb20e4
Reviewed-on: https://go-review.googlesource.com/22598
Reviewed-by: Austin Clements <austin@google.com>
Run-TryBot: Austin Clements <austin@google.com>
